Physics-Referenced and Domain-Decoupled Monitoring of Chronic Temperature–Humidity Sensor Drift in Grain Storage

We present a three-stage framework for monitoring chronic temperature–humidity sensor drift in grain storage. A target-masked coupled heat–moisture proper orthogonal decomposition (POD) surrogate produces an expectation independent of the target sensor; spatial-consistency and temperature–humidity-coupling evidence decouples genuine grain-condition changes from isolated sensor deviations; and a normalized three-objective physics-informed autoencoder trained with multi-gradient descent estimates signed drift, standardized severity, onset time, and a maintenance grade. The framework was evaluated on approximately 36 million records from 12 horizontal paddy-rice warehouses across three depots, including 127 labeled chronic-drift events and 20 reference-instrumented locations. Under Protocol A, which is the leave-one-depot-out protocol, the five-seed result reaches precision of 91.5%, recall of 92.9%, window PR-AUC of 0.908, mean/P95 detection delay of 7.2/12.3 days, and a grain-anomaly misattribution rate of 3.1%; the per-depot breakdown of this same protocol retains macro recall of 92.8±1.4%. On the 87 independently confirmed events, precision is 92.0% and the final sensor-drift identification rate is 90.8% (79/87; 97.5% of the 81 detected events). The independent physical reference, validated against the co-located reference instruments, reaches a bias-corrected RMSE of 1.15 ∘C and 0.56%RH; a representative 30-day window reduces RMSE from 0.689 to 0.294 ∘C and from 0.392 to 0.235%RH. The field-sensor aggregate (1.26 ∘C, 0.37%RH) is optimistic for humidity because the reference is calibrated to those same field sensors. Drift magnitude has an MAE of 0.22 ∘C/0.27%RH, onset a median error of 2.1 days, and the maintenance grade an agreement of 88.1% (37/42) against the reference-derived grade. Online latency is 16.2 ms on the normal monitoring path and 42.7 ms on the complete three-stage alarm path. The method separates source attribution from drift estimation, exposes weak drift before it exceeds sensor accuracy, and delivers maintenance-ready outputs at low online cost.